Memory System

MEMORY.md — Persistent Memory

The core of WindOp's memory is a single file: MEMORY.mdin your app data directory. It's loaded at the start of every session, so WindOp always knows your preferences and context.

Entries are organized into 4 categories:

## preference
- User prefers concise responses with code examples
- Always use TypeScript over JavaScript
- Prefer functional programming patterns

## knowledge
- Project uses Next.js 16 + Tailwind CSS 4 + Prisma
- Database is PostgreSQL on Supabase (self-hosted)
- Deployment via Docker on port 3000

## pattern
- When fixing CSS issues, check Tailwind v4 @theme directive first
- Database migrations go in prisma/migrations/ with timestamp prefix

## insight
- The auth module has a race condition on token refresh
- Build times spike when PostCSS processes large CSS files

WindOp writes to memory automatically when it learns something important. You can also ask it to remember specific things: "Remember that we use PostgreSQL, not MySQL".

Daily Logs

WindOp maintains timestamped daily logs that record what happened during each session. Logs are stored as markdown files organized by date: logs/YYYY-MM-DD.md.

Each entry has a type that categorizes what happened:

Daily logs help WindOp recall what happened in previous sessions. Ask "What did we work on yesterday?"and it'll check the logs.

Smart Context Management

WindOp doesn't dump all memory into every conversation — that would waste tokens and confuse the model. Instead, it uses semantic relevance scoring to load only the memory entries that matter for your current task.

• •Session start: MEMORY.md is loaded in full — it's always available
• •During conversation: Relevant daily log entries are pulled in when contextually relevant
• •Context pruning: When the conversation gets long, older messages are summarized while keeping key decisions and tool results
• •Auto-continuation: If the AI hits the output token limit mid-response, WindOp automatically continues from where it left off